When will quantum computing have its breakout moment?

Begin with what has actually been shown, because the gap between that and the capital flowing in is the episode’s subject. The idea traces to Feynman in the 1980s: simulating quantum phenomena on a classical machine can be intractable, taking something like the age of the universe, so build a system that operates in the quantum realm instead. He scopes the prize carefully: quantum is an accelerator offering enormous gains on a narrow class of applications that happen to matter to a lot of people, not a general-purpose replacement for classical computing. Four decades on, his description of the state of play is deflationary. Hardware exists that you can kick the tires on, not quite commercial. No vendor has demonstrated what he would call dramatic performance gains over a classical counterpart; the work is still toy problems. And the benchmarks that generate headlines are the specific thing he objects to. His example is boson sampling, which he explains as simulating a pachinko machine: tracking how a thousand balls deflect off every pin is classically intractable, and a quantum system can represent it naturally because photons branch probabilistically. It was run, years ago, and the people who ran it said it had no practical application. Worse, one organization claimed a speedup of ten to the twenty-eighth power on that kind of benchmark and concluded in a blog post that this proved quantum operates across multiple universes. Sorensen’s retort is that engineers designing aircraft and crash tests do not want to hear that their compute requires additional universes. So the class of results the money is responding to is a class he explicitly discounts.

Why the distance from a toy to a tool is years rather than months has two answers, and the second is the one he thinks is underrated. The hardware answer is the noisy intermediate-scale era, now ending. Qubits are error-prone enough that you do not run an algorithm once and read an answer; you run it a thousand times and hope the correct result dominates the histogram, appearing perhaps 70 or 80 percent of the time. Error correction converts error-prone physical qubits into error-free logical ones, at a cost: a decade ago it took thousands of physical qubits to make one logical qubit, and hardware, architecture and algorithmic advances have been driving that ratio down. The target is roughly a million physical qubits yielding perhaps thousands of logical ones, which is where he says real science starts, and company roadmaps put that three to five years out. The lesson he draws is that a raw qubit count is a poor indicator and the error-correction scheme is the better one. But then the software answer, which he thinks is the harder problem. Quantum algorithms are not intuitive to creatures too large to be quantum, and there is no deep corpus to draw on. His comparison is the Navier-Stokes equations, nearly two hundred years old and only useful once machines could run them; quantum has had a fraction of that time, and beyond Shor’s algorithm for breaking current encryption he says there have not been many landmark applications. The metaphor he borrows, and admires, is that everyone is building a car in their garage while the algorithms are the roads, and eventually somebody has to open the door and drive somewhere. He calls the algorithmic problem the greater hill to climb over the next decade and adds that it is underfunded by governments and academia. That decade sits awkwardly beside his three-to-four-year hardware date, and he does not reconcile the two.

Usefulness sorts into three classes. First, simulating quantum phenomena directly, where classical methods require so many simplifying assumptions the results cannot be fully trusted: battery materials, catalyst design including catalysts for cleaner oil and gas products, protein and molecule design, and computational chemistry. Second, optimization, which carries an important wrinkle. Optimization does not have to be perfect to be valuable, so a noisy machine returning a merely better answer is still worth using where a small improvement multiplies across a fleet or a workforce. Third and most speculatively, the standard computational kernels under most simulation software, such as finite element analysis and computational fluid dynamics, where he cites a Rolls-Royce experiment that beat what quantum algorithm theory predicted. What he likes there is the implication that practitioners with heuristics may outrun the theory, which is how high performance computing always advanced. Against AI for materials discovery, he lands on complement rather than substitute: AI is data-driven, bounded by what is known, opaque and non-reproducible, so a human stays in the loop, while quantum computes from the physics without needing a large corpus and offers explainability and reproducibility.

Which brings him to what he is actually worried about, and it is not the physics. Eighty-five organizations are trying to become quantum hardware suppliers, more than have ever supplied high performance computers, laptops or smartphones across the entire histories of those markets. He says flatly that 70 could fail within two years without damaging the sector, because consolidation is a necessary part of the journey. His fear is interpretation. When five take down rounds and ten fold, investors conclude something is wrong, governments that spent billions ask what they got, and end users decline to commit to a vendor that may not exist in three years. The trajectory of the technology would be unharmed; the perception would not. His evidence that the capital is undisciplined is a company he consulted for, cheap and proud of it, that received a check for several hundred million dollars and whose chief executive told him there was no plan for it. He adds that the investors he speaks to often do not understand the technology or its timelines.